Configure function choices

By registering your plugin classes with the kernel, all of their functions become available to the AI model. The model can then automatically choose from any of these functions based on the prompt. Use AddFromType to register all functions from a class as plugins.

using Microsoft.SemanticKernel;

IKernelBuilder builder = Kernel.CreateBuilder();
builder.AddOpenAIChatCompletion("<model-id>", "<api-key>");
builder.Plugins.AddFromType<WeatherForecastUtils>();
builder.Plugins.AddFromType<DateTimeUtils>();

Kernel kernel = builder.Build();

PromptExecutionSettings settings = new() { FunctionChoiceBehavior = FunctionChoiceBehavior.Auto() };

await kernel.InvokePromptAsync("What is the likely color of the sky in Boston?", new(settings));

When you register plugin classes with the kernel, all of their functions become available to the AI model. The model can then automatically choose from any of these functions based on the prompt. In Python, use add_plugin to register your plugin classes and their functions.

import os
from semantic_kernel import Kernel
from semantic_kernel.connectors.ai.open_ai import AzureChatCompletion, OpenAIPromptExecutionSettings
from semantic_kernel.functions.kernel_function_decorator import kernel_function
from semantic_kernel.contents.chat_history import ChatHistory

class WeatherForecastUtils:
    @kernel_function(name="GetWeatherForCity", description="Gets the weather for a given city.")
    def get_weather_for_city(self, city: str) -> str:
        return "Sunny"  # Stub for demo

class DateTimeUtils:
    @kernel_function(name="GetCurrentUtcDateTime", description="Gets the current UTC date and time.")
    def get_current_utc_date_time(self) -> str:
        import datetime
        return datetime.datetime.utcnow().isoformat()

deployment_name = os.getenv("DEPLOYMENT_NAME")
endpoint = os.getenv("PROJECT_ENDPOINT")
api_key = os.getenv("PROJECT_KEY")

kernel = Kernel()
chat_service = AzureChatCompletion(
    deployment_name=deployment_name,
    endpoint=endpoint,
    api_key=api_key
)
kernel.add_service(chat_service, "chat_completion")
kernel.add_plugin(WeatherForecastUtils(), "WeatherForecastUtils")
kernel.add_plugin(DateTimeUtils(), "DateTimeUtils")

settings = OpenAIPromptExecutionSettings()
chat_history = ChatHistory()
chat_history.add_user_message("What is the likely color of the sky in Boston?")

result = chat_service.get_chat_message_content(
    chat_history=chat_history,
    kernel=kernel,
    settings=settings
)
print(result)

Use Plugins.GetFunction to select specific functions to advertise to the model. You can then pass these to FunctionChoiceBehavior.Auto(functions: ...) in your settings.

KernelFunction getWeatherForCity = kernel.Plugins.GetFunction("WeatherForecastUtils", "GetWeatherForCity");
KernelFunction getCurrentTime = kernel.Plugins.GetFunction("DateTimeUtils", "GetCurrentUtcDateTime");

PromptExecutionSettings settings = new() { FunctionChoiceBehavior = FunctionChoiceBehavior.Auto(functions: [getWeatherForCity, getCurrentTime]) };

await kernel.InvokePromptAsync("What is the likely color of the sky in Boston?", new(settings));

You control which functions are available by only registering the plugins you want the model to access. The SDK doesn't currently support restricting advertised functions directly in settings.

# As of SK 1.31, Python SDK does not support restricting advertised functions directly in settings.
# The closest approach is to only register the plugins you want available.

kernel = Kernel()
chat_service = AzureChatCompletion(
    deployment_name=deployment_name,
    endpoint=endpoint,
    api_key=api_key
)
kernel.add_service(chat_service, "chat_completion")
kernel.add_plugin(WeatherForecastUtils(), "WeatherForecastUtils")
# Only add DateTimeUtils if you want it available:
# kernel.add_plugin(DateTimeUtils(), "DateTimeUtils")

settings = OpenAIPromptExecutionSettings()
chat_history = ChatHistory()
chat_history.add_user_message("What is the likely color of the sky in Boston?")

result = chat_service.get_chat_message_content(
    chat_history=chat_history,
    kernel=kernel,
    settings=settings
)
print(result)

Set the FunctionChoiceBehavior to use an empty list of functions to disable function calling.

PromptExecutionSettings settings = new() { FunctionChoiceBehavior = FunctionChoiceBehavior.Auto(functions: []) };

await kernel.InvokePromptAsync("What is the likely color of the sky in Boston?", new(settings));

In Python, don't register any plugins with the kernel to disable function calling.

# In Python, you can disable function calling by not registering any plugins,
# or by using settings that prevent tool/function use if supported in your SDK version.
# As of SK 1.31, you can use the following pattern:

settings = OpenAIPromptExecutionSettings()  # No plugins registered, so no functions available

kernel = Kernel()
chat_service = AzureChatCompletion(
    deployment_name=deployment_name,
    endpoint=endpoint,
    api_key=api_key
)
kernel.add_service(chat_service, "chat_completion")
# Do NOT add any plugins

chat_history = ChatHistory()
chat_history.add_user_message("What is the likely color of the sky in Boston?")

result = chat_service.get_chat_message_content(
    chat_history=chat_history,
    kernel=kernel,
    settings=settings
)
print(result)

Use the FunctionChoiceBehavior class to configure these behaviors in C#.

In Python, these function behaviors are controlled by which plugins you register and, in some cases, by prompt engineering. The SDK doesn't currently provide direct equivalents for all C# behaviors.